Core Concepts
Invertible Residual Rescaling Models (IRRM) achieve state-of-the-art performance in image rescaling tasks using a lightweight and efficient architecture.
Abstract
The paper proposes a novel Invertible Residual Rescaling Model (IRRM) for image rescaling, which aims to reconstruct high-resolution (HR) images from their low-resolution (LR) counterparts. The key contributions are:
IRRM introduces Residual Downscaling Modules (RDMs) with long skip connections, which allow the model to focus on learning high-frequency information while easing the flow of information. Each RDM contains several Invertible Residual Blocks (IRBs) with short connections to enhance the non-linear representation capability of the model.
The proposed IRRM outperforms previous state-of-the-art methods like IRN and HCFlow on various benchmark datasets, while using much fewer parameters and computations. Specifically, IRRM-M achieves comparable performance to IRN with only 1/4 of the parameters, and IRRM-S performs well beyond previous super-resolution methods with less than 1M parameters.
Extensive experiments demonstrate the excellent extensibility of IRRM. The model with residual connections and enhanced residual blocks (RB) can be scaled up to achieve better performance, while the model without residual connections suffers from unstable training and degraded performance as the model size increases.
The paper also analyzes the influence of the latent variable z on the reconstructed HR images, showing that IRRM is insensitive to the Gaussian distribution of z and can effectively preserve image details.
Stats
IRRM achieves PSNR gains of at least 0.3 dB over HCFlow and IRN in the ×4 rescaling while only using 60% parameters and 50% FLOPs.
IRRM-L achieves PSNR of 46.41 dB and SSIM of 0.9921 in the ×2 rescaling, outperforming previous state-of-the-art methods by a large margin.
Quotes
"Our proposed IRRM introduces the Invertible Residual Block (IRB), which incorporates short skip connections to enhance the model's nonlinear representational ability. This addition significantly improves the extensibility of the model."
"With long and short skip connections, abundant information can be bypassed and thus ease the flow of information."